【Keras】训练时显存out of memory的解决办法——fit_generator函数

问题描述：建立好model之后，用model.fit()函数进行训练，发现超出显存容量(一共有12G)
问题分析：fit()函数训练时，将全部训练集载入显存之后，才开始分批训练。显然很容易就超出了12G
解决办法：用fit_generator函数进行训练

fit_generator函数将训练集分批载入显存，但需要自定义其第一个参数——generator函数，从而分批将训练集送入显存

def data_generator(data, targets, batch_size):
    batches = (len(data) + batch_size - 1)//batch_size
     while(True):
         for i in range(batches):
              X = data[i*batch_size : (i+1)*batch_size]
              Y = targets[i*batch_size : (i+1)*batch_size]
              yield (X, Y)

调用fit_generator时的方法

model.fit_generator(generator = data_generator(X_train, Y_train, batch_size),
                    steps_per_epoch = (len(data) + batch_size - 1) // batch_size,
                    epochs = num_epochs,
                    verbose = 1,
                    callbacks = callbacks,
                    validation_data = (X_val, Y_val)
)

参考链接：https://zhuanlan.zhihu.com/p/23250782
http://blog.csdn.net/sinat_26917383/article/details/74922230